SiMa.ai Hardware Documentation

This guide covers hardware setup, firmware, and low-level interfaces for SiMa.ai's MLSoC Modalix DevKit products. It is the dedicated hardware reference — software, ML, and pipeline tooling are documented separately.

About the MLSoC Modalix

The MLSoC Modalix is SiMa.ai's second-generation machine-learning system-on-chip. It combines an Arm Cortex-A65 application complex, a high-throughput Machine Learning Accelerator (MLA), an Image Signal Processor (ISP) for camera ingest, and a Computer Vision Unit (CVU) for classical vision workloads — all on a single die.

This integration lets developers deploy multimodal, generative, and vision AI pipelines at the edge without stitching together discrete accelerators, host CPUs, and camera bridges. Every product documented here — DevKits, Early Access kits, and the PCIe card — is built around the same MLSoC Modalix silicon, so applications written for one form factor port cleanly to the others.

MLSoC Modalix block diagram.

DevKit Portfolio

SiMa.ai offers a portfolio of MLSoC Modalix-based development kits and products designed to support a broad spectrum of edge AI applications — from efficient vision inference to advanced multimodal and generative AI workloads. Click a kit below to jump to its detailed page, or compare features in the table.

Modalix DevKit

A compact, power-efficient module for embedding Modalix AI capabilities into custom hardware designs.

Modalix Early Access DevKit

Designed for modern AI at the edge with support for LLMs, multimodal inputs, and generative AI workloads.

Modalix Early Access PCIe Card

Half-Height/Half-Length (HHHL) card designed for extensible edge compute systems.

Feature Comparison

Feature	Modalix DevKit	Modalix Early Access DevKit	Modalix Early Access PCIe Card
Documentation & OS
Product Brief	view	view	—
Preloaded Operating System	eLxr Linux	eLxr Linux	eLxr Linux
Compute
ARM Cores	8x ARM Cortex-A65 @ 1.4GHz	8x ARM Cortex-A65 @ 1.4GHz	8x ARM Cortex-A65 @ 1.4GHz
ISP (Image Signal Processor)	ARM C-71 @ 1.2 GHz	ARM C-71 @ 1.2 GHz	ARM C-71 @ 1.2 GHz
CVU (Computer Vision Unit)	Synopsys EV74 @ 750 16-bit GOPS	Synopsys EV74 @ 750 16-bit GOPS	Synopsys EV74 @ 750 16-bit GOPS
Memory & Storage
RAM Size	32GB LPDDR5	64 GB LPDDR5	32GB LPDDR5
Storage	16GB eMMC	10 GB eMMC	16GB eMMC
SD Card Slot	✖	✔	✔
NVMe (PCIe)	✔ 500GB	✖	✖
Networking
Ethernet	1x 1GbE	1x 1GbE (end0), 1x 10GbE (end1), 2x 10GbE SFP+ (end2/3)	1x 1GbE
WiFi/LTE	2x M.2 slots via carrier board, pending s/w support	✖	✖
Camera Inputs
MIPI CSI	2x 2-lane MIPI CSI	4x 4-lane MIPI CSI	✖
GMSL2 over FAKRA	✖	✖	2x GMSL2 over FAKRA
I/O & Display
GPIO / Headers	40-pin GPIO header	40-pin GPIO header	✖
USB	4 USB 3.0 ports	✖	✖
HDMI	1 HDMI 1.4 port	✖	✖
Graphics Controller	Silicon Motion SM768	✖	✖
Video Codecs
H.264/H.265 Encoder	4kp60	4kp60	4kp60
MJPEG Encoder	4kp30	4kp30	4kp30
H.264/H.265 Decoder	4kp60	4kp60	4kp60
AV1 and MJPEG Decoder	4kp60	4kp60	4kp60
Form Factor
Use as a PCIe Card in a host	✖	✖	✔

Order Your DevKit

Deployment Architectures

Modern machine learning and inferencing applications demand flexible architectures to address a variety of deployment scenarios. SiMa.ai's solutions are designed to adapt to these needs, offering configurations that optimize performance, efficiency, and scalability. Whether integrated into a larger system or utilized as a standalone device, the MLSoC Modalix ensures seamless adaptability for diverse use cases.

MLSoC Modalix as a Standalone Device

In this architecture, the MLSoC Modalix operates independently as a self-contained device. It is particularly well-suited for applications where compactness, efficiency, and minimal power consumption are critical.

Key Use Cases

• Edge AI Applications: Deployed at the edge to perform inferencing without relying on a central server or cloud infrastructure. Ideal for applications like smart cameras, industrial IoT devices, or autonomous robots.
• Cost-Sensitive Deployments: Reduces the need for additional hardware, making it a cost-effective solution for standalone operations.
• Power-Constrained Environments: Optimized for scenarios where energy efficiency is paramount, such as remote monitoring systems powered by batteries or solar panels.

Advantages

• Self-Contained: Does not require a host system, simplifying deployment and reducing system complexity.
• Energy Efficient: Designed for low power consumption, making it suitable for power-sensitive environments.
• Compact and Portable: The small form factor allows it to be easily deployed in space-constrained scenarios.

Typical Data Flow

1. Data is received directly from network interfaces or sensors.
2. The MLSoC Modalix is loaded with a NEAT application that defines the on-device inference pipeline.
3. The MLSoC Modalix performs inferencing and processes the data locally.
4. Results are sent to other devices or systems via network connections for further action or visualization.

Set up Standalone Mode System Using DevKit

MLSoC Modalix as PCIe Card

In this architecture, the MLSoC Modalix functions as a PCIe card that integrates seamlessly into a host system. This setup is ideal for applications where the host system—such as a server, desktop, or workstation—requires additional computational resources for inferencing tasks but already has sufficient capabilities for data storage and processing.

Key Use Cases

• High-Performance Systems: Suitable for data centers or high-end workstations where large-scale data processing and storage are critical. The host system handles CPU intensive tasks, while the MLSoC Modalix performs dedicated inferencing operations.
• Extended I/O Solutions: Target system may require additional I/O capabilities that are not readily available on the MLSoC Modalix, such as USB interfaces or other sensory inputs. This architecture enables the system to leverage the host system's extended I/O resources for applications like video analytics or AI model inference at scale.

Advantages

• Enhanced Computational Resources: Augments the host system's capabilities by offloading inferencing tasks to the dedicated hardware.
• Flexibility: Ideal for environments where additional inferencing cards can be added as demand grows.

Typical Data Processing Flow

1. The host system captures data from sensors, peripherals, storage, or network interfaces.
2. The MLSoC Modalix is loaded with a NEAT application packaging the inference pipeline and the adapted AI model.
3. Data is sent to the MLSoC Modalix via the PCIe interface for inferencing.
4. Results are processed by the host system or forwarded to downstream systems for further action.

Set up a PCIe Mode System Using DevKit

Get Started

Pick the path that matches your setup. Each guide covers serial-console access, network bring-up, and firmware management for that mode.

Quick Start Guide

Step-by-step interactive walkthrough for first-time DevKit bring-up.

Standalone Mode

Run the MLSoC Modalix as a self-contained edge device — no host required.

PCIe Mode

Install the Modalix PCIe card into a host system for offloaded inference.